"Include a diagram" produced decks with no picture. Three separate faults, each
hiding the next, found by generating the same deck after fixing each one.
First, deck generation ran on the default 4000-token budget. A deck's JSON is
several times the size of the prose it holds, so a long deck came back truncated,
failed to parse, and fell back to markdown — which has no way to request a
figure, so the model described one instead and the slide rendered a literal
"![Placeholder: Flow diagram ...]" as its first bullet. Deck generation now gets
room, and the fallback says how the reply failed: empty, cut short at N
characters, or not a deck.
Second, the figure request sat inside the layout vocabulary, one line among
forty, and the model passed over it. It goes last now, after the author's own
instructions — the same placement lesson the image tool taught earlier.
Third, and the one that actually mattered: image_prompt is only read on the
figure and image types, so an image_prompt on a bullets slide was dropped in
silence. The instruction said "add image_prompt to N slides" without saying
which types carry one. It now names them, and a misplaced request is honoured
rather than discarded — a slide with words becomes a figure, one without becomes
a full-slide image.
Image markup is also stripped wherever text enters a slide, on both paths: a
described figure is not a figure, and a bullet of raw markdown is worse than no
bullet.
Verified end to end after: the same request produced a deck with one figure, the
job completed, and the exported pptx carries one embedded image across 21 slides.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Dv6sqaY6Vq3ChZHMem3cnU
A deck's JSON is several times the size of the prose it holds, and generation
used the default 4000-token budget — raised for refine and for slide review, but
never here. A sixteen-slide deck ran past it, came back truncated, failed to
parse, and fell back to markdown. Markdown has no way to ask for a figure, so
the model described one instead and the slide rendered a literal
"![Placeholder: Flow diagram — "Neonate with rash" → ...]" as its first bullet,
above the steps it was meant to illustrate. That is why no generated deck was
arriving with an image.
Deck generation now gets room for a deck. The fallback also says how the reply
failed — empty, cut short at N characters, or simply not a deck — because those
want different fixes and "not usable" covered all three.
Image markup is stripped wherever text enters a slide, on both the deck and
markdown paths, since a described figure is not a figure and a bullet of raw
markdown is worse than no bullet. The model is also told plainly: if a figure is
wanted say so with image_prompt, and if that is not on offer, write the slide
without one rather than describing the picture you would have drawn.
Separately, the Documentation list showed ARCHITECTURE, CLINICAL_ASSISTANT,
DEVELOPMENT, MODULE_CONVENTIONS and SCALING shouting in caps with underscores
intact: the label builder replaced hyphens but not underscores, and uppercased
the first letter of each word rather than normalising the case, so a
SHOUTING_FILENAME stayed shouting. It now reads "Clinical Assistant", keeps
acronyms as acronyms (AI, API, OpenID, LiteLLM) and leaves joining words lower.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Dv6sqaY6Vq3ChZHMem3cnU
savedFigureIds was declared inside the review branch, so with no reviewer
configured — the default, and what everyone is running — it was undefined by the
time the INSERT stringified it. JSON.stringify(undefined) is not a string, the
column is NOT NULL, and every generation failed with "Generation failed". `var`
is function-scoped, so nothing complained until the database did.
This is the second bug of exactly this shape in this file, so the test asserts
position rather than presence: the value must be declared before both the review
and the insert read it.
Found by the logging added in the same change, which is the other half of this
commit. Every modification now says what it did:
[my-resources] refine id=29 path=deck outcome=applied 13→14 slides changed=yes
[my-resources] refine id=37 path=markdown outcome=applied 2635→3018 chars changed=yes
CHANGED=no is warn-level and deliberately shouty, because that is the failure
worth catching: the response says success either way, the row updates, and the
download is identical — which is exactly how the deck bug went unnoticed. A
refusal logs its reason. ped_ai_resource_refine_total{path,outcome} counts the
same thing over time, so "did that modification do anything" is answerable
without watching logs live.
Verified across every path rather than the one that was broken: a deck
presentation modified and exported to both pptx and docx carries the change; a
legacy presentation with no stored deck still takes the markdown path and
carries it; an article generates, modifies and exports; and a presentation
generates with the reviewer off.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Dv6sqaY6Vq3ChZHMem3cnU
The shape vocabulary is deliberately small, which leaves the question of what to
add next. Rather than guess, it now records demand.
Two signals, because a model asks both ways. It can say so outright —
{"kind":"unsupported","need":"a SmartArt cycle of four stages"}, which draws
nothing and is told about in the same file that validates it — or it can reach
for a kind, chart type or slide type that does not exist, which is the more
common way of asking and just as much of a signal.
Both produce a log line naming what was wanted and the topic it came up on, and
increment ped_ai_deck_vocabulary_gap_total{wanted}, so it can be counted over
time in Grafana rather than noticed once and forgotten. Deduplicated per
generation and capped at twelve: a model that asks for a hundred things it cannot
have should not write a hundred log lines. It can never fail a generation — it is
a note to whoever decides what to build next.
This is also the answer to whether to run model-authored code in a sandbox
instead. The log will say whether the gap is real. Some of it is not closeable by
any sandbox, being python-pptx's own ceiling — no SmartArt, no animations or
transitions, limited chart types — and a sandbox would only let a model write
code against the same library and hit the same wall. Documented in
docs/my-resources.md, which the in-app Docs tab serves directly.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Dv6sqaY6Vq3ChZHMem3cnU
The nine layouts are a fixed vocabulary and a good default, but "lay the three
severity levels out left to right with arrows between them" had no expression in
them at all. A "custom" slide now carries a list of shapes: positioned text,
eight autoshape families, lines, images, tables, and native PowerPoint charts —
column, bar, line, pie, doughnut.
Coordinates are percentages of the slide rather than EMU, because a model
reasons about "the left half" and not about 12192000. Shapes draw in array
order, so a later one sits on top.
The model never emits Python. It names shapes and the renderer draws them:
running model-authored code to lay out a slide would be an enormous amount of
trust to buy a feature, on a server holding clinical data and secrets.
Validation lives beside the text that teaches the vocabulary, in one file, so
what the model is told about is exactly what is accepted. Kinds are an
allowlist, colours must be six hex digits, coordinates are clamped inside the
slide — a shape at x=95 w=30 is cut to the edge rather than drawn half off it —
counts are capped, a pie is held to one series, and anything that cannot be
understood is dropped. A custom slide that loses every shape becomes a plain one
rather than a heading over an empty frame, and one bad shape is caught in the
renderer so it cannot cost the slide it sits on.
A figure on a custom slide is requested through an image shape, drawn by the
same path as any other, and attached by job id. Word renders a custom slide as
its words in reading order with its tables and figures — lossy, and better than
dropping the slide.
Verified live end to end: asked to "lay the three severity levels out left to
right as coloured boxes with arrows between them", the model produced
[rect arrow rect arrow rect], chose green/amber/red itself, and the rendered
slide was looked at. A column chart beside its commentary renders with real axes
and gridlines.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Dv6sqaY6Vq3ChZHMem3cnU
Pandoc reads markdown, so every Word export had to flatten the resource to
markdown first — and a deck flattened to markdown stops being one. A comparison
became two headings and two lists, a callout became bold text, and a figure
became nothing at all, because markdown has nowhere to put it.
src/utils/docSpec.js reduces either source to the same blocks: a stored deck
where there is one, the markdown where there is not. scripts/render_docx.py
draws them. A comparison comes out as a labelled two-column table, a callout as
a shaded box, a table as a real table, a figure embedded at its own aspect ratio
with its caption, and speaker notes as muted indented text.
The deck wins over the markdown beside it, because that markdown is a
serialisation of the deck and reading it instead would be reading a lossy copy of
what is right there.
Word now carries the figures too. The export route skipped fetching them for
docx, which was correct when pandoc could not place them and wrong the moment
this could.
Pandoc stays installed and stays the fallback: a plainer document beats a failed
download. Both renderers now share one spawn helper.
Verified end to end: a deck with two figures exported as a six-page Word document
with both images embedded (537KB, two files in word/media), rendered to PDF and
looked at — the comparison is a labelled table, the figure sits at its true
aspect ratio, and the notes read as notes.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Dv6sqaY6Vq3ChZHMem3cnU
The model that writes a deck never sees it. It cannot tell that slide four
overflowed, that a nine-item list would read better in two columns, or that two
labelled groups want to be a comparison — those are facts about the rendered
page, not about the text. So each generated deck is now rendered to PDF through
Gotenberg, rasterised to one image per slide with pdftoppm, and shown to a
vision model.
Off unless an administrator names a reviewer, in its own admin card because it
is the one setting that spends money on every generation without a user having
asked for anything. One pass, on generation only: a second pass costs as much as
the first and fixes far less, and refining is a text edit.
It returns a patch, not a deck. Asking for the corrected deck back put the reply
in proportion to the deck rather than to the number of problems, and a
fourteen-slide deck came back cut off mid-object at every output budget the
provider would honour — measured twice before changing shape.
The patch is better for a second reason. The reviewer names a slide and an
action — two columns, one column, split after bullet N, compare with these two
labels — and the server moves the text it already has. The words never pass
through the model, so a review cannot reword, drop or invent a single bullet.
That is a stronger guarantee than instructing it not to and checking afterwards.
The check runs anyway, because a bug in applyChanges would be as bad as a model
rewriting the words and worse for being trusted: body text must come out the
same multiset, figures the same set, and a heading may only be reused or
extended. A continuation heading is the reviewer's one piece of text and is
replaced when it does not continue anything.
Nothing here can fail a generation — no reviewer, an unreachable one, an
unparseable reply, a deck too long to look at, or a patch that applies to
nothing each return the deck that was written.
Verified end to end against a deck with a deliberately overloaded slide: three
slides rendered and sent, one change returned, ten bullets split into five and
five under "Stepwise Management … (continued)", text intact. Left switched off;
enable it under Admin.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Dv6sqaY6Vq3ChZHMem3cnU
Nothing documented My Resources, the slide renderer, PubMed or web search, and
the authentication doc predated both sign-in codes and registration invitations.
docs/my-resources.md is new and covers the feature end to end: what a resource
is, where its material comes from, why both searches run in the route rather
than as tools the model never called, why keyword engines get the topic while
retrieval gets the instruction too, how a presentation is designed as a deck
rather than written as markdown, the separate multi-image path, and what the
export pipeline is made of.
docs/authentication.md gains sign-in codes — storage, lifetime, reuse,
supersession, guessing, and that two-factor still applies — and registration
invitations, including the exact condition that decides when a code may be
deleted and why it is written to match the status the list displays. Both new
rate limits are in the table, with a note that Express matches app.use paths on
segment boundaries, so a new sign-in endpoint needs its own limiter or it has
none at all.
docs/deployment.md now says what the runtime image carries and why — pandoc for
Word, python3 with apk-installed lxml and pillow for the slide renderer,
python-pptx pinned, and that PDF conversion is not in the image at all but goes
to Gotenberg, so Word and PowerPoint still work when it is down.
docs/configuration.md picks up LOGIN_RATE_LIMIT_MAX, LOGIN_CODE_RATE_LIMIT_MAX
and GOTENBERG_URL, none of which were listed. README gains a My Resources
section and indexes the two new docs.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Dv6sqaY6Vq3ChZHMem3cnU
An expired code is as dead as a used one and just as accounted for, so it is now
deletable. The rule the code enforces is the one that matters: a code that could
still be redeemed is never deleted, because that takes it off the list without
taking it out of anybody's inbox — the holder keeps something that looks valid,
it quietly stops working, and nothing is left to say who had it.
One condition, shared by the single delete and the bulk clear:
(used_at IS NOT NULL OR (revoked_at IS NULL AND expires_at <= NOW()))
Written that way rather than as "used OR past its date" because the second form
also catches a revoked code whose date has since passed — a row the list still
labels revoked and offers no delete on, so the button and the query would have
disagreed about the same row.
Revoked codes keep their rows. Revoking records a decision somebody took, and a
handful of them is not the clutter a pile of expired codes is.
Verified against the live database across every state: active refused, used
deleted, expired deleted, revoked refused, and revoked-with-a-past-date refused
rather than slipping through as expired.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Dv6sqaY6Vq3ChZHMem3cnU
The delete button was offered on every invitation regardless of state, and the
query behind it deleted any row it was given. Deleting an unused code takes it
off the list without taking it out of anybody's inbox: the person still holds
something that looks like a valid invitation, it silently stops working, and
there is no longer a record of who it went to or why. Revoke is what stops a
live code — it leaves the row behind, marked.
So the delete is now for spent codes only, in three places rather than one: the
query carries AND used_at IS NOT NULL, the route answers 409 with the reason
instead of pretending the row is missing, and the button is rendered only on a
used row.
A "Clear N used" control alongside, since the complaint was clutter and clearing
them one at a time is not much of an answer. Same rule — nothing unused or
revoked is touched — and it confirms first, because it is still a delete.
The bulk route is declared before /invites/:id, or Express reads "used" as an id.
Verified against the live database: deleting an unused invitation is refused and
the row survives, deleting a used one works, the bulk clear removes only used
ones, and the unused probe row was still there afterwards.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Dv6sqaY6Vq3ChZHMem3cnU
The sign-in screen asks for an email first, then offers both ways in together:
a six-digit code sent to that address, or the password. Beside rather than
instead — a code depends on mail being delivered and a password does not, so
neither may be the only route. "Use a different email" goes back a step, and
creating an account stays where it was.
What keeps it from being a second, weaker front door:
- Only a bcrypt hash is stored, so a code read out of the database is not a
working credential.
- Ten minutes, single use, marked used before the session is issued so a replay
cannot race it, and requesting a new one deletes the old.
- Five wrong guesses burn it. Six digits is a million possibilities, which is
plenty against a person and nothing against a script with unlimited tries.
- Requesting a code answers identically whether or not the address exists, and
every verify failure returns one message. A sign-in screen that says "no such
account" is a way of finding out who has one.
- Two-factor still applies: a code proves you can read the mailbox, which is one
factor, and an account that asked for a second still wants it.
- Its own rate limits, tighter for requesting than for attempting, because
requesting sends mail to someone else's address. These had to be separate
limiters: Express matches app.use paths on segment boundaries, so
/api/auth/login does not cover /api/auth/login-code — checked against a real
router rather than assumed.
Two bugs found while building it, both mine:
authFetch keeps an allowlist of endpoints callable with no verified owner and
rejects everything else before it is sent. The new endpoints were not on it, so
the request never left the browser and surfaced as "Connection error".
reveal() hid elements by appending 'hidden' to className and showed them with a
non-global replace, so hiding twice left two copies and showing stripped one.
The "use a different email" link never reappeared. It uses classList now, which
is idempotent.
Verified against the running server: correct code signs in, the same code again
is refused, a superseded code is refused, five wrong guesses burn it, an expired
one is refused, and the stored value is a hash. In the browser: requesting a
code advances the screen, a wrong code is refused without losing the screen, and
the password route still signs in.
Not yet demonstrated: a correct code typed into the browser. The harness keeps
racing the one-live-code rule — the page's own request supersedes whatever code
the test holds, and with SMTP off the delivered one cannot be read. The same
request reaches the server on the wrong-code path, and the endpoint itself is
verified, but that last step is untested end to end.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Dv6sqaY6Vq3ChZHMem3cnU
Markdown could express about five of the things the renderer can draw, so the
model had no way to say "put this figure beside these three bullets" or "make
this a comparison with two labelled columns" — my parser inferred a layout from
the shape of a list, and inferring is what made every deck look the same.
A presentation is now described as a deck: the model returns JSON naming a
layout per slide and the prompt it wants each figure drawn from. Four layouts
were added to the renderer for it — two tinted labelled columns for a
comparison, a callout card for a red flag or a dose, a figure beside its
bullets, and a full-slide figure. Articles stay markdown, which is what prose
wants.
Markdown is still produced, serialised from the deck, so Word export and text
editing keep working and the stored artifact stays readable by a person. The
deck is stored alongside it because that serialisation is lossy by design:
round-tripping through markdown would throw away exactly the layout choices this
was built to capture. A resource made before this, or an article forced into
slides, still renders by inferring from its markdown.
Nothing here can cost more than the thing that went wrong. A reply that is not a
deck falls back to asking for markdown rather than saving the model's apology; a
malformed slide degrades to bullets rather than throwing; a comparison with one
column is not a comparison; a figure that cannot be queued leaves a slide of
text rather than an empty frame; and JSON wrapped in fences or a covering
sentence is read rather than refused.
Verified live on "croup versus epiglottitis": the model chose section, bullets,
table, compare, figure, callout and image layouts across thirteen slides, and
the exported deck was rendered to PDF, rasterised and looked at — the comparison
renders as two tinted cards, the red flag as a callout, and the figure sits
beside its bullets.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Dv6sqaY6Vq3ChZHMem3cnU
Pandoc's pptx writer was the ceiling on how good a generated deck could be, and
the model on top made no difference to it. It maps markdown onto a handful of
reference layouts with no per-slide layout, no positioning and no control over
how large an image is drawn, which is why every deck came out as bullets on a
template — and why autofit had to be injected into its emitted OOXML by hand
afterwards, because LibreOffice ignores the autofit pandoc leaves off.
scripts/render_pptx.py draws the deck and src/utils/slideSpec.js decides what
each slide is. Markdown stays the stored artifact, so "change slide 4" is still
a text edit and Word export is untouched — pandoc still writes docx, where its
output is good.
What that buys, all of it visible in a rendered deck rather than argued for:
- 16:9, not pandoc's 4:3.
- A pipe table becomes a real table with a header band and banded rows, not
eight lines of text with pipes in them.
- A list longer than seven items becomes two columns instead of a wall of text.
- Text is measured and sized to fit before the file is written, so nothing
depends on a renderer honouring autofit.
- Wrapped lines hang under the text instead of running back to the margin,
which is the clearest single tell that a deck was generated.
- An image is drawn at its own aspect ratio, centred, with a caption.
Figures now reach the deck at all, which they never did. They were queued and
shown on the page, but nothing recorded that they belonged to the resource, so
an export could not include them: user_resources.image_ids holds them, a
modification adds to that list rather than replacing it, and export fetches the
finished ones to a scratch directory. They are spread through the deck rather
than appended, because ending on three unexplained pictures is worse than
showing each near its material, and a References slide stays last.
If the renderer fails for any reason, pandoc still produces a deck — a plainer
deck beats a failed download.
Verified end to end: a seven-slide request with three figures exported as a
13-page deck; the slides were rendered to PDF, rasterised and looked at. All
three formats still download.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Dv6sqaY6Vq3ChZHMem3cnU
My Resources used imageTool.dispatch, which permits exactly one image per
request — "Only one image tool invocation is permitted per request". That is the
right rule for a chat reply and the wrong one for a twelve-slide deck where the
author asked for three figures. Rather than relax a limit the clinical assistant
and the Learning Hub also depend on, this adds a separate dispatcher for this
feature. Same queue, same storage, same my_resources workflow, same asset
endpoint — only the number of figures differs, bounded at six because each one
is a paid request.
"Use 3 diagrams" in the instructions is read as the number it is, and the
illustration option now says several are possible rather than promising one.
Three things had to be got right, each found by measuring rather than assuming:
The illustration guidance has to be the last thing in the prompt. Placed before
the output rules it lost — with the tool offered and the paragraph present, the
model returned 3297 characters of markdown and zero tool calls, while the same
tool and wording in a shorter prompt produced three calls.
Even last, it loses to a prompt carrying thirty library excerpts:
deterministically, with the library off "use 3 diagrams" made three calls and
with the library on it made none and wrote a longer deck instead. So when the
author names a number the call is required rather than merely offered. With no
number named the choice stays the model's.
And a model that has just made three tool calls tends to sign off instead of
writing: "I'll create the presentation and the three teaching diagrams." was
returned as the resource — 61 characters — because only a completely empty body
counted as missing. A body with no title block and no heading is not a Pandoc
document whatever its length, and if the continuation is no better than the
first attempt, whichever actually reads like a resource is kept.
Verified end to end with the library on: generate produced three figures and an
eight-slide deck; modify added two more figures and a ninth slide. The figures
were fetched and looked at — labelled airway anatomy, and a croup/epiglottitis/
bacterial tracheitis comparison.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Dv6sqaY6Vq3ChZHMem3cnU
The screen had been trimmed to controls with no explanation of what it was for;
"Private to you" said who could see it, not what it did. It now opens with a
sentence that says what you get and how you get it out, and the header reads
"Only you can see these".
The four choices — clinical library, PubMed, web, illustration — are one "Draw
on" group instead of four separate rows, with the library ticked by default.
Each option hides itself when an administrator has not enabled it, so nothing
appears that a person could tick and then be refused.
Modify offers exactly the same choices. It had none, so "add what the 2024 trial
showed" was answered from the model's memory rather than by looking anything up.
Generate and Modify now go through one gatherSources(), so they cannot drift
into offering different things or searching them differently.
Writing "include a diagram of the airway" in the instructions now switches the
illustration option on and says why, rather than the request being dropped in
silence. Switching it off by hand sticks — the hint then reminds instead of
fighting — and when no image model is configured it says so rather than
pretending. Both the generate and modify boxes behave this way.
Two things found by testing this rather than assuming it.
PubMed ANDs every mapped term, so one unrecognised word takes the query to zero.
"febrile seizures" returns six results and "febrile seizures in under-fives"
returns none; "the anatomy of croup: subglottic narrowing and the steeple sign"
returned none until it was narrowed to "anatomy croup", which returns six. A
query that finds nothing is now retried against progressively shorter versions
of itself, longest first, and the response says which query actually worked so
the screen cannot report one that found nothing. Those extra calls tripped
NCBI's three-a-second limit and produced a 429, so retries are spaced and the
first attempt waits for nothing. Separately, the searches run on the topic while
the library retrieval also gets the instruction: retrieval is semantic and
benefits from the context, but a keyword engine handed a whole sentence returns
nothing.
And when a search was asked for and came back empty, the prompt now says not to
invent a citation, a PMID or a URL to fill the gap. Without that the model
supplies them from memory, and a fabricated PMID looks exactly like a real one.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Dv6sqaY6Vq3ChZHMem3cnU
PubMed joins web search as an optional source for a generated resource: a
literature search on the topic, with abstracts, cited by PMID in References.
Off by default, admin-enabled, with its own optional API key (NCBI raises the
rate limit from 3/sec to 10/sec; it works without one).
Neither search is a tool any more, and that is the point. Offering them as
function calls meant the model decided whether to search, and with a prompt
ending "Output ONLY Pandoc markdown" it decided not to — every time, with and
without corpus grounding, no matter how the tool description was worded.
Calling callAI with the tool directly produced a correct pubmed_search call, so
the plumbing was never the problem. The search only ever needed the topic, and
the route knows the topic before it calls the model, so both searches now run up
front and their results go into the prompt as findings, exactly the way corpus
excerpts do. Ticking the box now means the search happened.
Verified live against deepseek-v4-flash: 30 corpus excerpts and 6 PubMed
results, and a References slide carrying both the library sources and four real
PMIDs (29562151, 38506440, 35721052, 28814254).
Three fixes to illustration, which had never once fired:
- The dispatch call had been lost in a refactor. The tool was still offered, the
model still called it, and the call was dropped, so no job was ever enqueued.
- imageContext was passed as a bare topic string where dispatch expects
{ request, history }, which made the bound request undefined.
- The prompt never mentioned the tool existed while explicitly demanding only
markdown — the same suppression that killed the searches. It now says an
illustration is available and that calling it is not a violation of that rule.
my_resources is its own image workflow rather than a reuse of learning_hub,
because generated_image_links only accepts learning_hub assets, and that is
exactly the barrier that keeps a private illustration out of published content.
The illustration renders in the panel, rather than a toast pointing at an image
history this feature does not have.
Verified end to end: job queued, rendered, and the asset served to its owner as
a correctly labelled subglottic-anatomy teaching diagram.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Dv6sqaY6Vq3ChZHMem3cnU
The one feature here that sends text outside the building, so the defaults are
the careful ones: disabled unless an administrator turns it on, opt-in per
generation even then, and the option is hidden entirely rather than shown as
something a user can tick and be refused.
Only the search query leaves. Library excerpts, the generated resource and
anything about the user never do. Both screens say so plainly, because a topic
typed while drafting clinical material can carry clinical detail and the
provider keeps its own logs.
Four providers behind one shape, so swapping changes nothing downstream: Tavily,
Serper over Google, Brave, and SearXNG — the only one where the query does not
reach a commercial third party at all, which is why it is worth supporting even
though it needs somewhere to run.
The tool description says when NOT to search, because a model handed a search
tool will reach for it constantly: not for settled clinical knowledge, which is
what the indexed library is for, and one search per resource. That last one is
enforced in the route with toolChoice: 'none' on the continuation rather than
trusted to the model.
A failed search never fails a generation — same contract as corpus retrieval.
The resource is written without it and the response says what was searched for
and what came back, so a query that left the network is visible rather than
silent.
The API key is masked on read and preserved when the field is left blank, the
handling the OIDC client secret already gets, so changing provider cannot
silently wipe a working key.
Verified on the running instance: with nothing configured, webSearchAvailable is
false, and a request asking for it anyway is ignored rather than honoured.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Dv6sqaY6Vq3ChZHMem3cnU
A generated deck was losing content. pandoc writes a bare <a:bodyPr/> on every
shape, which leaves the body with no autofit even though the slide master has
one, so a slide with too much on it is cut off mid-sentence and the rest is not
rendered at all. Reproduced and counted: eight bullets went in, three came out,
the third ending mid-word.
Every generated deck now carries <a:normAutofit/> on its body placeholders. No
fontScale, deliberately — the renderer works out the reduction, so a slide that
already fits is untouched, where a fixed scale would shrink all of them. The
same eight bullets now fit with nothing in the bottom 6% of the slide.
This is a floor, not a licence to overcrowd. The prompt still asks for one idea
per slide; this stops a long one becoming unreadable.
Also: an article is no longer offered as PowerPoint. A deck of paragraphs is not
a presentation. Word and PDF suit either kind, and the route refuses the
combination rather than relying on the button being absent.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Dv6sqaY6Vq3ChZHMem3cnU
Learning is moderator-owned: content published into categories that everyone
sees, behind router.use(moderatorMiddleware). That is right for institutional
material and wrong as the only way in — an ordinary user could not generate
anything at all.
So this is a separate pathway rather than a loosening of that one. Learning is
untouched; the moderator gate stays exactly where it was. A signed-in user can
generate a deck or an article for their own use, keep it, refine it and export
it, and nobody else ever sees it.
Private by construction. Every statement filters on the owner and there is no
route that returns another person's work, which a test asserts statement by
statement rather than trusting. The foreign key cascades, so deleting an account
takes its drafts with it. There is no category, no publish state and no sharing:
adding sharing later should be a deliberate feature, not something that leaks
out of a forgotten WHERE clause.
Markdown is the artifact. Every format is rendered from it on demand — pptx and
docx by pandoc, both carrying the house reference deck, and PDF by Gotenberg,
whose LibreOffice preserves a deck's layout in a way rendering from markdown
would not. That is what makes "add a slide on when to admit" a text edit rather
than a binary patch.
Gotenberg was published on the host but on a network of its own, so reaching it
from a container went out and back through the host gateway. It now joins
danvics_convert, owned by danvics-net like the others. PDF is the one export
allowed to fail: if that service is down, the deck and the document still
download and the error says which.
Verified end to end as a plain user: the moderator route still refuses with 403,
generation returned a deck grounded on 12 corpus excerpts, the library lists
only their own, pptx/docx/pdf all downloaded valid, "add a slide on when to
admit" put the slide in the right place and left References last, and an
unauthenticated request gets 401 while someone else's id gets 404.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Dv6sqaY6Vq3ChZHMem3cnU
Learning generated everything from the model alone. A deck on bronchiolitis was
whatever the model remembered about bronchiolitis, with no connection to the
documents this institution actually indexed — while the assistant had been
searching that corpus all along.
Same collection, deliberately. mcp_bge_m3_1024 is already embedded with
openrouter-bge-m3 at 1024 dimensions; a second index over the same documents
with the same embedder would be a copy that drifts. What differs is the budget:
a chat answer wants a few tight excerpts because the reader is waiting, a
teaching resource synthesises a whole topic. So learning.search_limit and
learning.context_chars default to 30 and 2500 against the assistant's 8 and
1400, and are separate keys so tuning one cannot move the other.
Not unbounded, though. "No limit" only moves the ceiling from a setting to the
model's context window, where overflow truncates the middle of the prompt
silently — the worst place to lose source material. 60 results and 8000
characters per excerpt are the caps.
Opt in per generation: a resource on something the library does not cover is
better written without it than padded with the nearest unrelated excerpts.
Retrieval never fails a generation — the resource is then written from the model
alone, which is what happened before this existed — and every response reports
what it was grounded on, so a caller can say "24 excerpts" or "the library had
nothing on this" rather than quietly serving ungrounded material.
Verified against the live corpus: bronchiolitis, neonatal jaundice and febrile
seizure each returned 12 excerpts and ~23k characters from Nelson, Rudolph and
the Pediatric Clinical Practice Guidelines. A deck generated through the full
chain came back with textbook specificity that is not general recall —
bronchiolar diameter, birth-weight thresholds, the full pathogen list.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Dv6sqaY6Vq3ChZHMem3cnU
npm audit reports 0 vulnerabilities. It reported 2 high and 2 moderate this
morning.
@google-cloud/vertexai was the last source of findings — gaxios and a uuid with
a missing buffer bounds check, neither reachable in this deployment because
GOOGLE_VERTEX_PROJECT is unset and the require sits inside that check. Dormant
is not the same as gone, and the provider is available through the gateway
anyway, so the direct path has been removed rather than left to rot:
- the SDK client and callVertex, which without the package could never run
- the dispatch and discovery branches that reached them
- VERTEX_MODELS, a list of ids nothing could route any more, and the two
places in adminConfig that concatenated it into the built-in set
- the health endpoint's vertex line, and the env vars documented for it
AI_PROVIDER=vertex now says where to configure the model instead of quietly
becoming something else. The Google STT and TTS paths keyed off the same
variable are untouched; neither ever used this SDK.
Verified after deploy: provider litellm, the assistant answers with 8 sources,
/api/models returns 10, and @aws-sdk/s3-request-presigner — which documents.js
needs for presigned MinIO URLs — is still declared and resolvable.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Dv6sqaY6Vq3ChZHMem3cnU
Measured in a real browser against the app rather than reasoned about.
The transcript boxes are contenteditable divs, and an editable div zooms on
focus exactly like an <input>. The earlier 16px sweep covered input, textarea
and select, so every workspace tab still zoomed while the calculators did not —
which is exactly what was reported. Every focusable text control in every tab
now measures 16px at phone width; the count of ones below it is zero.
Three ways a recording could end with nothing to show for it:
- Safari supports none of the audio/webm types and throws NotSupportedError
when handed one. Six modules built their own recorder on resume with
"opus, else audio/webm", so resuming threw there and the recording stopped.
There is now one codec chain in the app, and no module constructs a
MediaRecorder of its own.
- audio-recorder-failed is dispatched on document, and the encounter tab
stopped its recording on any of them. The assistant's microphone failing
ended a consultation being recorded in another tab. The recorder now
travels with the event and the listener checks it is its own.
- The server answers {success:true, text:''} for silence, and five modules
assigned that straight into the transcript — emptying the box the browser
had been filling live. It reads as a recording that vanished. Text is now
required before overwriting, and a recording that captured nothing says so
instead of resetting the button over an empty box.
Also: the citation counters were registered on prom-client's default registry
while the app serves its own, so they were never scraped. They read zero at
/metrics now instead of being absent, which is what the Grafana panels need.
And the reference linter passes for the first time, so scripts/e2e.sh gets past
its preflight: KaTeX is vendored (it was referenced by the assistant's LaTeX
rendering but never shipped — three 404s a page load and no math), and the
JavaScript left behind by the removed image picker, saved-chats toggle, image
gallery and visual-output panel is gone.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Dv6sqaY6Vq3ChZHMem3cnU
Citation quality
- A citation naming a source that never came back is never rendered as a
link, so it appears as plain text and nobody learns it happened. It is now
measured on the server, where the answer and the sources both exist, so it
is seen whether or not a browser rendered it.
- Four Prometheus counters feed a Grafana dashboard (Ped-AI Citation
Quality): answers, citations written, answers affected, and individual
unresolved markers. Only answers with at least one unresolved citation are
stored, with the question and the titles retrieval returned, so an operator
can judge whether retrieval came back thin or the model over-cited. Rows
expire after 30 days: this is a quality signal, not a transcript log.
- Both answer paths are covered. /chat/stream is normal; /chat is the
fallback the client uses when streaming fails, so auditing only the first
would have hidden exactly the answers produced under failure.
- The tracker is resolved on demand and allowed to be absent. Seven test
files load this route with a hand-built list of permitted imports, and
adding a hard dependency would mean editing all seven — and the eighth
written later would break. Observation must never be able to fail an
answer, so a missing module simply means no tracking.
- Metric registration reuses an already-registered counter, because this
module can legitimately load twice in one process.
SSO settings on mobile
- Six rows were laid out inline: flex with a 160px label and an input that
would not shrink, so on a phone the row was wider than the screen with
nothing to scroll and no way to reach the rest. They use .admin-row now,
which already stacks below 640px. Verified at 390px and 360px: nothing
off-screen, no sideways overflow.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Dv6sqaY6Vq3ChZHMem3cnU
With more than one admin, every setting in the panel was editable by all of
them — prompts, model policy, retrieval budgets, SMTP, email templates.
ADMIN_LOCKDOWN separates running the service from changing how it behaves.
It is an environment variable on purpose: a setting could be switched off by
the very admin it restrains, so lifting this needs host access and a restart.
The server is the control. One gate refuses configuration writes rather than
a check in each of the fifteen write routes, because that list grows and a
route added later would quietly miss it. Reads always pass — lockdown hides
nothing. Day-to-day operation stays available: invitations, announcements,
registration, feature flags, and the test endpoints, which persist nothing.
A setting invented later is locked until someone deliberately makes it
editable, rather than defaulting to open.
The panel disables what it cannot save and says why, but that is courtesy;
the refusal is what enforces it.
Two things this taught me, both fixed: my first version painted the panel
from an IIFE, which the module conventions forbid, and fetched the whole
config a second time just to read one flag — breaking the test that pins
admin loaders firing exactly once. The state now rides on the invites
response the panel already requests.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Dv6sqaY6Vq3ChZHMem3cnU
registration_enabled was a single switch: open to anyone, or closed to
everyone. This adds the setting an operator actually wants in between — open
to people you invited.
A code is single-use, expires (7 days by default, 90 maximum), and can be
revoked or deleted. It is stored hashed with only its last four characters
kept, because an invite grants account creation and a database dump should
not hand someone a working one. The code is readable exactly once, in the
response that creates it.
The claim is a single conditional UPDATE carrying every condition, so two
registrations racing the same code cannot both succeed. It happens after the
account exists, so a code is never spent on a failed registration — and if
the race is lost, the just-created account is removed rather than left behind
as a free registration. The rejection never says which of the four reasons
applied; distinguishing them would tell someone probing codes which guesses
were closer.
Codes avoid I, L, O and U so they survive being read aloud or copied off a
screen, and matching ignores case and separators.
The sign-up field appears only when the server says a code is required. The
admin card creates, lists, revokes and deletes, and carries the toggle.
Verified against the live database: create, claim, second claim refused,
unknown code refused, revoking a used code refused, delete. 684 tests pass.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Dv6sqaY6Vq3ChZHMem3cnU
The nc_semantic_search alias is gone from the MCP server, so accepting it
here would point retrieval at a tool that no longer exists. A stale
override now stops the app at startup instead of silently retrieving
nothing.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Dv6sqaY6Vq3ChZHMem3cnU
Three S3 configurations had grown separately — S3_* for documents,
GENERATED_IMAGES_S3_* for images, and AUDIO_BACKUPS_S3_* after them — with
different key names and their own client construction. That is why moving
storage meant hunting through several files.
src/utils/objectStorage.js now resolves settings for any purpose: its own
variables first, then the shared S3_* ones, with a per-purpose bucket name
(S3_BUCKET_AUDIO_BACKUPS). One endpoint plus three bucket names is enough
for the whole app, and a purpose that needs its own account still overrides
everything. Audio backups and documents use it; generated images keeps its
own tested storage module, whose variable names the resolver already
understands.
Nothing existing has to change: S3_ACCESS_KEY_ID, S3_SECRET_ACCESS_KEY and
the AWS_* fallbacks still resolve, and path-style addressing keeps each
purpose's previous default — off for documents, so a Backblaze endpoint
behaves as before, on where a custom endpoint implies MinIO. A _FILE
credential now always beats an inline one, so a mounted secret cannot be
shadowed by an inherited environment variable.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Dv6sqaY6Vq3ChZHMem3cnU
Recording
- A screen wake lock is held for as long as a recording runs. Browsers drop
the lock whenever the page is hidden, so it is taken again on return —
without that, one glance away ended it for the session. The lock is
reference counted (two recorders cannot release each other's), never
requested while hidden (the request would just be rejected), and a denial
or an unsupported browser leaves the recording running.
- Signing out releases it and stops the recording; nothing is sent, because
the session that owned the audio is gone.
- start() on an already-running recorder is now a no-op instead of replacing
the MediaRecorder and silently dropping everything captured so far.
- A recording that ends by itself — recorder error, or the microphone taken
by another app, unplugged or revoked — takes the same path as pressing
Stop, so it is transcribed and stored rather than left in a tab that still
says "recording". Moving around the workspace already kept recording.
Retention
- Every recording is kept for 24 hours now, not only the ones whose
transcription failed. /api/transcribe already has the audio, so this costs
no second upload, and a storage failure is logged rather than thrown: it
must never lose the transcription someone is waiting for.
- One store (src/utils/audioBackupStore.js) is shared by /api/transcribe and
/api/audio-backups so the two cannot drift. Payload goes to object storage
when AUDIO_BACKUPS_S3_* is set and to the encrypted Postgres column
otherwise; metadata always stays in Postgres, so listing, ownership and
expiry behave the same either way. Object keys are scoped by owner, and
the expiry sweep deletes the object with the row.
Verified against the live database: round trip byte-identical, another user
reads null, 950 -> 48 bytes compressed, expired rows take their objects.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Dv6sqaY6Vq3ChZHMem3cnU
Security
- nodemailer 9.0.1 -> 9.1.1, clearing four high advisories, two of which
are delivery bugs that matter for an app that sends mail: recipient-domain
validation bypass via RFC 5322 comments, and an IDN/punycode allow-list
bypass, both of which can route mail to an attacker-controlled domain.
Live transcription
- The Settings picker was a hardcoded list of six ids
(local-whisper-*, local-parakeet-v3, gemini-*). None of them resolve on
this gateway, and /api/transcribe prefers the user's choice over the admin
default, so picking one broke every recording with "Invalid model name".
Verified against the live gateway: local-whisper-large-v3-turbo -> 400.
- The picker now lists what /model/info advertises as audio_transcription,
cached for five minutes, with the built-in list kept only as a fallback
and the admin default marked.
- The pipeline itself is healthy: local-kokoro-tts produced 92KB of speech
and mistral-voxtral-mini-transcribe returned the sentence back verbatim.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Dv6sqaY6Vq3ChZHMem3cnU
Feature audit
- Settings claimed you could reference a template by saying "use my normal
physical exam" in dictation. No phrase handling exists anywhere, and the
prompt says the opposite: "Never copy clinical content from a template —
only formatting and structure." So a template can never supply findings.
The text now says what happens, and keeps the true privacy statement that
only template categories go to the AI (Custom is filtered out in
/memories/context by AI_CONTEXT_CATEGORIES).
- Templates themselves are real: CRUD plus /memories/context, injected as
style hints by hpi, soap, sickVisit, wellVisit, edEncounters and
hospitalCourse, behind the `memories` feature flag.
Docs
- docs/CLINICAL_ASSISTANT.md listed six settings and offered `deepl`, which
no longer exists in the code. The table now covers all seventeen keys the
server reads, with their fallbacks, plus how a model reaches a user.
Testing
- Every test file's stdout is now pure TAP, which is the stream node:test
parses results from. Three sources: a leftover debug console.log dumping
600 characters of HTML, page modules logging into a JSDOM without a
virtual console, and the server startup banners. The banners are guarded
by NODE_TEST_CONTEXT, set only inside node:test children, so production
and `node server.js` output is unchanged (verified both ways).
- Three consecutive full-suite runs at 671/671.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Dv6sqaY6Vq3ChZHMem3cnU
Gallery tiles are 56px but were downloading the full ~280kB original. Previews
are now rendered with sharp and stored beside the originals in the same MinIO
bucket under a thumbs/ prefix, so nothing about credentials, lifecycle or backup
changes. Measured on live assets: 216-294kB originals become 13-19kB at 256px,
about 16x smaller; 640px is about 4x.
Both paths, as asked:
- Rendered when a job completes, so the first viewer never waits for a resize.
A preview failure never unmakes a finished job.
- Rendered on demand for anything that has none — the existing 26 images work
immediately with no backfill required, and the result is stored for next time.
Boundaries that matter more than the speed:
- Only 256 and 640 are honoured. An open width parameter would let a caller
drive arbitrary resizes.
- Permission is checked against the ORIGINAL before a preview is served, so a
preview can never widen who can see an image.
- Previews carry their own SHA-256 and owner headers, because the client
verifies both on every asset; sending the original's checksum would be
rejected as tampering, which is that check working correctly.
- Still private, no-store. The client asset pattern was widened to exactly
?w=256 and ?w=640 and nothing else.
Client-side downscaling stays as the fallback when a preview cannot be produced.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_018e1PLqrKgAM9jQhFKRnbLd
Sources (correcting what I built earlier)
The previous toggle branched the SYSTEM PROMPT, so the same question could get a
different answer depending on a display setting — the bias this was meant to
avoid. The prompt is now unconditional: buildSystemPrompt takes no display
argument and is byte-identical either way. Hiding sources happens on the way out
— the server omits them and strips the now-orphaned [n] markers from the copy it
sends. The answer is generated, stored and exported with citations intact, so
turning the setting back on restores them without re-asking anything. Renamed to
clinical_assistant.show_sources; the old key is still honoured.
Signed-out preview (admin opt-in, default off)
A visitor may try the assistant; reaching for the workspace asks them to sign in.
Deliberately narrow:
- Reachable paths are an exact allow-list, not a pattern, so a new endpoint is
private unless someone adds it on purpose.
- A preview visitor gets no identity at all (id: null), so nothing can be owned,
saved, billed or addressed to them.
- The image tool is withheld rather than left to fail on a null owner, and no
audit rows are written.
- A caller presenting a token is authenticated normally, so preview can never
downgrade a real session; if the setting cannot be read, authentication is
required.
- Actions needing an account are hidden rather than offered and refused.
Composer
The bar above the transcript is gone. Patient take home, Export PDF, Download
transcript and Attach images moved into a + menu in the composer, and the model
selector moved beside send — shown only when there is more than one model, as
before. Both views now start at the same top edge, so switching modes cannot
nudge the page up or down. On an empty transcript the tiled ground runs behind
and below the composer, which floats on it above centre.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01GmpYHPSLGmXGZMyLpn2Lbe
Multimodal removal
The multimodal path called nc_multimodal_search against a second hardcoded
collection whose embedding service (multimodal-embeddings:7999) was never
deployed and ENABLE_MULTIMODAL_RAG has always been false, so it only ever logged
"multimodal search skipped". Removed rather than left as dead weight:
- clinicalRetrieval: normalizeMcpMultimodalResponse, isVisualSourceQuery,
isRadiologyQuery, buildMultimodalSearchQuery, classifyAndRerankMultimodalResults,
selectMultimodalResults, visualIntent, visualMetadataScore,
shouldRejectVisualSource, allowsFrontMatterQuery, looksLikeFrontMatterPage,
looksLikeTextOnlyPage and MULTIMODAL_CANDIDATE_LIMIT (~140 lines).
- clinicalMcpClient: multimodalSearch.
- The route's visual/text slot split is gone; the whole search limit is text.
- The "[visual PDF page match]" prompt label and the "visual PDF page" source
badge are gone with it.
Adding models
Model availability could only be ticked from what the gateway advertised, so an
admin could never offer a model discovery did not list. Each list now has a text
field: a typed id joins the same checkbox list, is enabled by default, is
de-duplicated, and persists through the normal allowed_models save.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01BkfrkQwA4YGrGw9LZSpeAq
New setting clinical_assistant.citations_enabled (default true, admin checkbox).
With it off, retrieval, grounding and every other rule are unchanged — answers
are still built only from retrieved sources — but:
- buildSystemPrompt swaps only the citation block: the "cite factual claims with
[1]" rules are replaced with "do not include citations, source numbers or
bracketed markers", and the note that the sourcing requirement itself is
unchanged. Grounding, scope, table formatting and tone rules are byte-identical
between the two modes.
- The server strips any stray [n] the model emits anyway, from the stored answer
rather than only the view, so saved chats and exports match what was shown.
- No sources are sent to the client at all, and the status endpoint reports the
mode so the UI hides the Sources panel and gives its 330px column back to the
chat instead of showing an empty rail.
Validated as a boolean in adminConfig, like the feature.* keys.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01BkfrkQwA4YGrGw9LZSpeAq
Mobile
body.assistant-workspace .assistant-layout is both later in the file and more
specific than the .assistant-layout rule inside @media (max-width:640px), so
phones were inheriting height: calc(100vh - 64px) — an offset for the app header
that workspace mode already hides — plus grid rules on a flex container. The
topbar then overlapped the content. Added a matching-specificity mobile override
restoring 100dvh and the flex layout.
Workspace menu in the rail
The saved-chats rail gains a collapsible Workspace section listing the app's own
tabs, so the assistant sidebar carries both its chats and the workspace
navigation. It is built from the real .tab-btn elements rather than a second copy
of the menu, so tabs added, renamed or hidden in index.html follow automatically
and admin-only tabs stay hidden.
Prompt
Rewritten as principles instead of an enumerated rulebook: "respond to the user's
latest message, not to an earlier one", "if it carries no question, ask what they
would like you to look up", "never repeat a previous answer", and for the tool
"use it when the latest message asks for a picture, or for a change to one you
just made, and not otherwise". No example words in any language remain — the
model reads the message as written. 1032 -> 788 characters.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01BkfrkQwA4YGrGw9LZSpeAq
The previous fix put a hand-maintained multilingual word list in the route. That
does not generalise — the bug report itself was "Окей", and the next one would be
a language not on the list. The model already has the tool, the system prompt and
the whole conversation, so the policy belongs where it reads it.
Removed from src/routes/clinicalAssistant.js:
- GREETING_RE and its short-circuit (an ASCII keyword list that decided, before
the model ever saw the message, that a greeting deserved a canned reply).
- The IMAGE_NOUN / IMAGE_REPEAT vocabulary lists and the answer-repetition
backstop added earlier today.
Moved into the prompt and the tool definition:
- DEFAULT_BEHAVIOR now says that a greeting, acknowledgement or thanks in ANY
language gets a one-sentence request for a clinical question, and that a
previous turn's answer must never be repeated to justify a second image.
- The generate_image description says to call it ONLY when the user's latest
message asks for a picture or a change to one just made, and that an
acknowledgement of an existing image is not such a request.
Both are admin-editable (clinical_assistant.system_behavior), so this can now be
tuned without a deploy.
dispatchImageRequestFallback and IMAGE_REQUEST_PATTERN stay: that is the
compatibility path for a serving model that writes the image prompt as text
instead of calling the tool. It reads the USER's message only, never the model's
answer, so it cannot replay a previous turn — it was not the cause of this bug.
The DEFAULT_BEHAVIOR byte-hash lock in prompt-administration.test.js is updated
deliberately, which is what that guard is for.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01BkfrkQwA4YGrGw9LZSpeAq
Translation
- Stop scrubbing markdown before sending it to LibreTranslate. The scrub
deleted ordered-list numbering ("1. Give amoxicillin" -> "Give amoxicillin"),
flattened tables into ambiguous whitespace and ate underscores inside
identifiers. Raw markdown now goes to the translator unchanged.
- Render the translation through the same markdown pipeline as the original
bubble, with the message's own sources, so [n] markers come back as the usual
clickable .assistant-cite chips instead of escaped literal text. Headings,
lists and tables survive with them.
- When the translator drops citation markers, surface the affected sources in a
recovery block rather than letting the evidence disappear.
- Image cards are live nodes: they are now re-attached on every path out of a
translation (success, failure and Show original), so a failed translation no
longer silently removes a generating image from the message.
Patient take home
- Add a language selector to the take-home modal, reusing the existing
/translate endpoint and offering only what the local LibreTranslate reports.
- Copy, Export and Email carry what the caregiver is actually reading; the
original stays canonical behind "Original".
Conversation budget
- The admin field no longer prefills with the environment value, which turned
the next Save into an accidental override and made the documented "leave
empty to use the environment" path unreachable. The effective limit is shown
as a placeholder instead.
- Report source 'default' honestly instead of naming an unset env var.
- The load-failure notice now lands on the <p> instead of an <input>'s
textContent, where it rendered nothing.
- One validator for the budget everywhere: conversationLimit() replaces a
parseInt that accepted "120000abc".
Other
- /assistant is addressed by its URL, not by ped_last_tab, so "/" no longer
reopens the assistant; the URL follows tab changes and Back leaves it.
- Remove the dead DeepL path (it referenced an undefined DEEPL_BASES) and stop
offering admins a provider the server silently ignores.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01BkfrkQwA4YGrGw9LZSpeAq
